Tao Lu

Researcher · 3D Vision & Embodied AI

Tao Lu鲁涛

I study how machines perceive, reconstruct, and act in the 3D world.

These days I work on embodied world models — teaching robots from human video, rebuilding real scenes as interactive gyms, and world action models that predict before they act. Earlier, I worked on point clouds and LiDAR during my PhD at Nanjing University, then on 3D Gaussian Splatting (Scaffold-GS, Octree-GS, GSDF) at Shanghai AI Lab and as a postdoc at Brown University.

Scholar GitHub CV

New Real2Gym, InternW0-Δ, InfiniHand and GeoVerse are out on arXiv.

Portrait of Tao Lu
fig. 1 — me, in Gaussians

01 Research

From points to worlds.

One thread runs through my work: giving machines a 3D understanding of the world that is good enough to act on.

  1. move to orbit

    012018 — 2023

    Perceive

    Learning directly from point clouds and LiDAR — efficient point backbones, large kernels at linear cost, and camera–LiDAR fusion for flow and detection.

    CGA-Net CamLiFlow LinK SparseBEV

  2. hover for anchors

    022023 — now

    Reconstruct

    Structured 3D Gaussians for real-time, view-adaptive rendering — anchors, levels of detail, SDF coupling, and feed-forward splatting from unposed images.

    Scaffold-GS Octree-GS GSDF AnySplat

  3. click to set a goal

    032025 — now

    Act

    World models that predict, then act — robot skills from egocentric human video, real-to-sim gyms built from videos, and large-scale world action models.

    InternW0-Δ Real2Gym Skel-WAM InfiniHand

02 Publications

Papers

* equal contribution  ·  † corresponding
Full list on Google Scholar

03 Contact

Say hello.

Always happy to talk about 3D vision, world models and robot learning — or about working together.